Compositionality
The meaning of a whole is determined by its parts and how they combine — the principle that lets a finite vocabulary express unbounded thoughts, and the standard AI systems are measured against.
Reviewed July 16, 2026Stable
When not to use it
- As a verdict on whether a model understands. Compositional behaviour is measurable; understanding is a separate and contested question.
- On benchmarks where the novel combinations appear in the prompt, which tests pattern application rather than generalisation.
Reach for something else instead
- Decomposition prompting, which substitutes an explicit process for compositional machinery the model does not reliably deploy.
- Symbolic components for the parts of a task where recombination must be guaranteed.
- Coverage-based evaluation, accepting that the training distribution may be dense enough that recombination rarely arises.
Sources & further reading
- Fodor & Pylyshyn (1988), Connectionism and Cognitive Architecture — the systematicity challenge.
- Kim & Linzen (2020), COGS: A Compositional Generalization Challenge — the in-distribution versus structural gap.
- Lake & Baroni (2023), Human-like systematic generalization through a meta-learning neural network — systematicity as a property of training rather than architecture.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Testing out-of-distribution by varying subject matter, which holds structure fixed and misses this entirely.
- Reading strong benchmark scores as compositional competence when the benchmark drew test items from the training distribution.
- Treating the question as settled in either direction. The evidence has moved substantially and is confounded.
At a glance
FieldFoundations
OriginFrege, formalised by Fodor & Pylyshyn 1988
Measured bystructural recombination
Typical gap96-99% in-distribution against 16-35% out
Elicitableby decomposition prompting
DifficultyTechnical
Where this sits
A destination. 4 concepts lead here, and nothing in the corpus depends on it.
4Levelsteps in
4Needs firstconcepts
0Opens upnothing further
1Areastays here
Learn these firstGeneralization
Computed from the prerequisite graph, not assigned. How this works